scale range for neural network
1 Ansicht (letzte 30 Tage)
Ältere Kommentare anzeigen
mustafa alnasser
am 24 Okt. 2014
Beantwortet: Greg Heath
am 27 Okt. 2014
Dear All;
I have in neural network, : My input range is starting with small value , then it become big ( 1000 times of the small values) which a lot of data accumulate in small range and make conflict between detail in this range , how can we solve it . Also, if I have data range overlap at the boundary which may lead to misinterpretation, what is the best way to overcome this issue.
Regards;
0 Kommentare
Akzeptierte Antwort
Greg Heath
am 27 Okt. 2014
I don't have any specific code. I would experiment with the following.
1. Worry about the gating later.
2. First, determine, by trial and error, subsets of effective ranges for inputs of fitnet. This will not be an easy task. It is quite possible that nonlinear input transformations (e.g., logs or powers ) may be be helpful.
3. Since fitnet defaults to mapminmax transformations of inputs and outputs before other calculations, what has to be determined is how to choose the different ranges of inputs that will be transformed to [-1,1].
4. If there are c input range categories the gating net targets should be {0,1} c-dimensional unit vectors. The transformations between the vectors and category indices are
targets = ind2vec(indices)
indices = vec2ind(targets)
Hope this helps.
Thank you for formally accepting my answer
Greg
0 Kommentare
Weitere Antworten (1)
Greg Heath
am 25 Okt. 2014
This is easily solved by using a gating net that sends the input to a following net designed for a specific range of inputs.
Hope this helps
Thank you for formally accepting my answer
Greg
Siehe auch
Kategorien
Mehr zu Sequence and Numeric Feature Data Workflows finden Sie in Help Center und File Exchange
Community Treasure Hunt
Find the treasures in MATLAB Central and discover how the community can help you!
Start Hunting!